CODEWITHSUNDEEP
sqlapache-sparkaggregate-functionsapache-spark-sqlpyspark
lilloraffa
lilloraffa

Reputation: 1397

SparkSQL: conditional sum using two columns

I hope you can help me with this. I have a DF as follows:

val df = sc.parallelize(Seq(
  (1, "a", "2014-12-01", "2015-01-01", 100), 
  (2, "a", "2014-12-01", "2015-01-02", 150),
  (3, "a", "2014-12-01", "2015-01-03", 120), 
  (4, "b", "2015-12-15", "2015-01-01", 100)
)).toDF("id", "prodId", "dateIns", "dateTrans", "value")
.withColumn("dateIns", to_date($"dateIns")
.withColumn("dateTrans", to_date($"dateTrans"))

I would love to do a groupBy prodId and aggregate 'value' summing it for ranges of dates defined by the difference between the column 'dateIns' and 'dateTrans'. In particular, I would like to have a way to define a conditional sum that sums all values within a predefined max difference between the above mentioned columns. I.e. all value that happened between 10, 20, 30 days from dateIns ('dateTrans' - 'dateIns' <=10, 20, 30).

Is there any predefined aggregated function in spark that allows doing conditional sums? Do you recommend develop a aggr. UDF (if so, any suggestions)? I'm using pySpqrk, but very happy to get Scala solutions as well. Thanks a lot!

Upvotes: 3

Views: 11377

Answers (1)

zero323
zero323

Reputation: 330073

Lets make your a little bit more interesting so there are some events in the window:

val df = sc.parallelize(Seq(
  (1, "a", "2014-12-30", "2015-01-01", 100), 
  (2, "a", "2014-12-21", "2015-01-02", 150),
  (3, "a", "2014-12-10", "2015-01-03", 120), 
  (4, "b", "2014-12-05", "2015-01-01", 100)
)).toDF("id", "prodId", "dateIns", "dateTrans", "value")
.withColumn("dateIns", to_date($"dateIns"))
.withColumn("dateTrans", to_date($"dateTrans"))

What you need is more or less something like this:

import org.apache.spark.sql.functions.{col, datediff, lit, sum}

// Find difference in tens of days 
val diff = (datediff(col("dateTrans"), col("dateIns")) / 10)
  .cast("integer") * 10

val dfWithDiff = df.withColumn("diff", diff)

val aggregated = dfWithDiff 
  .where((col("diff") < 30) && (col("diff") >= 0))
  .groupBy(col("prodId"), col("diff"))
  .agg(sum(col("value")))

And the results

aggregated.show
// +------+----+----------+
// |prodId|diff|sum(value)|
// +------+----+----------+
// |     a|  20|       120|
// |     b|  20|       100|
// |     a|   0|       100|
// |     a|  10|       150|
// +------+----+----------+

where diff is a lower bound for the range (0 -> [0, 10), 10 -> [10, 20), ...). This will work in PySpark as well if you remove val and adjust imports.

Edit (aggregate per column):

val exprs = Seq(0, 10,  20).map(x => sum(
  when(col("diff") === lit(x), col("value"))
    .otherwise(lit(0)))
    .alias(x.toString))

dfWithDiff.groupBy(col("prodId")).agg(exprs.head, exprs.tail: _*).show

// +------+---+---+---+
// |prodId|  0| 10| 20|
// +------+---+---+---+
// |     a|100|150|120|
// |     b|  0|  0|100|
// +------+---+---+---+

with Python equivalent:

from pyspark.sql.functions import *

def make_col(x):
   cnd = when(col("diff") == lit(x), col("value")).otherwise(lit(0))
   return sum(cnd).alias(str(x))

exprs = [make_col(x) for x in range(0, 30, 10)]
dfWithDiff.groupBy(col("prodId")).agg(*exprs).show()   

## +------+---+---+---+
## |prodId|  0| 10| 20|
## +------+---+---+---+
## |     a|100|150|120|
## |     b|  0|  0|100|
## +------+---+---+---+

Upvotes: 10

Related Questions